BACK
A leading Swiss insurance provider set out to modernize one of its most manual and error-prone processes: collecting and verifying vehicle registration documents for internal fleets and private insurance customers. Existing market OCR solutions were too slow, too expensive, or too inflexible to meet the company’s performance and cost goals. Friendly Nerds was brought in to design a custom solution using an event driven AI and serverless architecture that would drastically reduce processing time and operational costs, while delivering a smooth and secure user experience.
To insure vehicles for both internal fleets and external insurance customers, users were required to submit a vehicle registration document and manually enter the necessary data, which was slow and error prone. The insurer explored a third-party solution based on visual recognition and AI, but it proved too expensive and slow, taking up to 50 seconds per scan. The business needed a faster, more efficient system that could process data reliably without manual effort and scale automatically to control costs. The solution also had to handle poorly aligned or folded images and ensure secure, high-confidence data for insurance policy creation.
Friendly Nerds developed a fully serverless and event-driven platform that enables customers to upload photos of their vehicle registration document through an app or chatbot. The system tolerates poor alignment or folded documents and uses OCR to extract relevant data automatically.
Each document is passed through a series of validation checks, including image quality scoring and plausibility of extracted values. A confidence score determines whether AI-based automation is sufficient or if fallback to human verification is needed. This dual-layer approach ensures high accuracy without sacrificing speed.
By combining AWS Lambda, Step Functions, S3, and API Gateway with AI and business logic, the platform remains cost-effective at scale and delivers an end-to-end experience in under five seconds. The solution was implemented iteratively, adding features and test cases over multiple short development cycles.